Turnkey Rules

The 22 rules split by plane: one per protected data category on the retrieval side, one per action-risk class on the action side. The set exists so the product is useful before your first policy meeting.

The set is deliberately small. 22 rules a security engineer can actually read replaced an earlier generation of more than 60 near-duplicates, one per outcome variant.

Each rule crosses two agent attributes: business function, the need-to-know axis, and trust tier, a deliberate human decision that AI never sets. A finance agent at the highest tier reads a full SSN; an engineering code-review agent sees it masked; a low-trust agent in the same function gets masking or denial.

The encoding is fail-closed. An agent with an unknown function or an unset tier lands in a protective branch that masks or denies by default.

Credentials copied into docs, tokens in runbooks, keys in archived notes: most estates have them. The default catalog catches these recognizable high-risk shapes before any classification project completes.

The floor is honestly scoped. It targets recognizable classes and value shapes; it does not claim semantic understanding of everything sensitive in your domain, and it does not replace classifying your own material.

Rules are text you can read before anything enforces. New and edited rules are simulated against recent real traffic before they take effect, so a rule that would break legitimate work is caught before it ships.

The rollout follows the traffic. Agents start in monitor mode with every would-be decision recorded. Uncovered actions surface with a recommended covering rule attached, drafts are created disabled for a human to enable, and the no-coverage default tightens class by class as your coverage numbers justify it.